DyFram’s R score is an engine-generated assessment output derived from the relationships among a system’s classified dimensions. It is not simply the product’s category, the model’s reputation, or the governance level the system will ultimately receive.
The distinction is essential. Classification describes the deployment. The R score expresses the engine’s assessment under defined inputs and calibration. Jurisdictional thresholds then translate that assessment into a base governance level. Sensitive answer patterns and jurisdictional rules can add further consequences.
An R score is useful only when readers understand what was assessed and which versioned process produced it. A precise-looking number without a boundary can create more confidence than knowledge.
Why relationships matter
The same capability can have different consequences depending on its use and environment. Reading a document for a private summary differs from using that reading to rank applicants. A tool that edits a fictional record differs from one that edits a real entitlement record. Human review can meaningfully constrain an action, or it can exist only as a nominal checkpoint.
These are relationships between dimensions. Capability does not tell the whole story without use. Potential harm does not tell the whole story without scale, reversibility, and context. Human involvement does not establish protection unless its authority and workflow position are understood.
DyFram’s Multi-Dimensional Categorization Matrix, or MDCM, records seven dimensions: Capability Profile, Use Case, IMPACT, ACTOR, artificial intelligence (AI) Ownership Class, CONTROL, and CONTEXT. Their scores and metadata form the DyFram Classification Signature. The engine uses dimensional relationships to yield R.
For a builder, the useful explanation begins with those relationships. A reviewer should be able to connect a consequential permission, a use, or an oversight arrangement to the assessment it helps produce. That explanatory chain matters more than presenting a precise-looking result without showing the deployment facts behind it.
What the score is not
R is not a probability of harm unless a validated interpretation explicitly establishes that meaning. It should not be described as “a 40 percent chance of harm” because the score happens to be 40. It is also not automatically a measurement in degrees exposure, written °exp. This article treats R and °exp separately rather than assuming a conversion between them.
R is not a governance level. A jurisdiction’s primary-sector thresholds are required to obtain the base G-level. Nor is R permission. An allowed or prohibited posture under the Conditional Prohibition Protocol cannot be reduced to whether a score falls below an ordinary governance threshold.
Surface DyFram does not weight model identity because reliable model weights are not claimed. Actual changes to deployment characteristics can still alter the assessment. A model replacement that changes enabled capability or effective control should not be ignored merely because model names are excluded from weighting.
Illustrative example: the same drafting function, different action paths
Imagine two hypothetical insurance-support tools. Both draft explanations of a claim file. The first operates on synthetic training records and cannot communicate externally. The second reads live personal records, produces a recommended disposition, and places that recommendation into a workflow where staff usually accept it.
The text-generation function is similar, but the systems differ in data posture, institutional reliance, affected interests, and control. The assessment should reflect those differences rather than treating “drafting” as a complete description.
Now suppose the second tool adds automatic closure of certain claims. That expands the action path and may change the meaning of human involvement. Reassessment should examine the actual configuration, not assume the old R remains valid because the user interface looks similar.
The example shows why the engine needs dimensional relationships rather than a label-based shortcut. The live information, institutional reliance, and ability to close a claim are different facts from the ability to draft text. An assessment should preserve that distinction and make its effect explainable.
How R becomes a base governance level
After R is produced, it is compared with the relevant jurisdiction’s primary-sector threshold profile. The result is a base level from G1 through G7. Those levels describe standardized governance intensity.
A different jurisdiction can choose a different threshold profile for the same primary sector. A comparable R can therefore produce different base levels without the engine contradicting itself. The difference may be a policy sensitivity, not a disagreement about the classified system.
Red-band pattern checks are another mechanism. If a matched band requires a higher governance level, the final level rises. If the base level is already higher, it remains higher and the band’s specific obligations are still added. This is why a single R number cannot communicate every governance consequence.
The government-controlled Jurisdictional Adaptation Layer is fetched before the governance obligations are assembled. The final package should make the assessment-to-obligation chain traceable rather than printing a score beside an unexplained list.
Reproducibility and evidence
The same inputs should reproduce the same outcome when the relevant questionnaire, engine, calibration, and threshold profiles are held constant. A comparison across versions requires examining what changed. Two similarly formatted scores are not necessarily comparable measurements.
The quality of inputs remains decisive. If a builder declares read-only access while the system can actually write, reproducibility can faithfully reproduce a flawed assessment. Consistency is not the same as truth.
The useful promise of R is clarity: an assessment tied to a defined system, with a path back to the characteristics that shaped it. Preserve scope and versions, test consequential assumptions, and keep policy thresholds distinct from the assessment. The number then becomes a starting point for a precise conversation, not the end of one.
Key takeaways
- R is an assessment output derived from dimensional relationships, not a product label or model ranking.
- R, G-level, permission posture, and °exp should not be treated as interchangeable quantities.
- Jurisdictional primary-sector thresholds translate R into a base governance level; red-band patterns add separate consequences.
- Comparability requires consistent scope, inputs, versions, and calibration—not merely matching number formats.
- Reproducible calculations can still rest on inaccurate declarations, so evidence remains necessary.